Iris open-weight search agents hit 88.6 on BrowseComp, weights on HF
AllSpark's Qwen-based models top open deep-research at 35B and 397B, run on any OpenAI-compatible endpoint, and show context management is the real lever.
Copy markdown
The scores: near-frontier, and fully open
Iris-pro (397B, 17B active) posts 88.6 on BrowseComp, 92.9 on DeepSearchQA, and 56.4 on HLE; Iris-mini (35B, 3B active) hits 82.2 / 86.9 / 52.3. AllSpark calls them the strongest open-source search agents in each size class, and the weights are CC-BY 4.0 with a 256K context.
The real lever is context management, not scale
The authors report that inference-time context management is worth more on these benchmarks than most reported gaps between systems. That's a retraining-free win you can port straight into your own ReAct agent loop.
Build this weekend: a private deep-research agent
Weights and the full Iris-Harness — agent loop, browsing tools, context manager, and graders — are on GitHub and Hugging Face now, and run against any OpenAI-compatible endpoint. Point vLLM at Iris-mini and you have a local agent that browses, reasons, and cites with no API bill.
The catch: the training recipe lands later
The models and eval harness ship today, but AllSpark says the data-construction and training pipelines are still 'coming soon.' You can run and fine-tune Iris now; reproducing its SFT-RL 'climbing' training from scratch has to wait.